Speech feature compensation in multiple model based speech recognition system using vts-based environmental parameter estimation

摘要

Multiple-model based speech recognition (MMSR) has been shown to be quite successful in noisy speech recognition. In this study, we propose a method to improve recognition performance by mitigating the mismatch in noise/channel type for an MMSR solution. We propose a novel method to reduce the effect of noise and channel mismatch by compensating the test noisy speech in the log-spectrum domain. We derive the relation between the log-spectrum vectors in the test and training noisy speech by using vector Taylor series (VTS) algorithm. Based on it, minimum mean square error estimation of the training log-spectrum vectors is obtained from the test noisy vectors by iteratively estimating environmental parameters. The estimated training vectors are used for recognition to reduce the noise and channel mismatch. We could find that the proposed method achieved WER reduction based on the Aurora2 task by +18.7% compared with a conventional MMSR method.